Andrew Caballero-Reynolds/AFP/Getty Images
Chat GPT-5, the next version of OpenAI's language model that changed everything, is on the way. Business Insider previously reported that some customers have received demos of the new version, and speculation about improvements is rife. The stakes are high for this model as it will solidify OpenAI's place at the forefront of AI or shove it to the middle of the pack.
As AI developers wait to test GPT-5, some have found a new way to bootstrap their way to more advanced AI without an entirely new model.
Jared Quincy Davis is the CEO of Foundry, an AI cloud service provider employing new strategies to make graphics processing unit computing more efficient.
The startup emerged from stealth in March with $80 million in seed funding co-led by Lightspeed Venture Partners and Sequoia Capital, for which Davis was formerly a startup scout specializing in AI, and a $350 million valuation.
Foundry's strategies center on cycling AI workloads across GPUs to maximize GPU utilization. Fewer idle GPUs then translate to more affordable computing, as Davis explained on the No Priors podcast released Thursday.
Davis' method of boosting the performance of existing large language models, which fits within an emerging classification of "Compound AI Systems," was explained in a research paper published last month.
The gist is that if you ask a model like Chat GPT-4 the same question repeatedly, you can use another, much smaller model to analyze those responses and pick the best one.
"More and more often, to go beyond the capabilities and frontier accessible to today's state of the art models and kind of get GPT-5 or GPT-6 early, practitioners are starting to do these things, oftentimes implicitly, where they'll call the current state-of-the-art model many, many times," Davis told podcast hosts Sarah Guo and Elad Gil.
Davis explained that his repeated "calling" of models is what some of the smartest, yet smallest models on the market have done in their training process. But right now, this method is only effective for certain kinds of queries — essentially questions with an answer that is easier to check than generate.
"For the kind of subjects you would expect, math, physics, electrical engineering, this type of approach was really helpful," Davis said.
In his test, Davis gave GPT-4 a challenging math problem. The model only got the correct answer 4% of the time, but could tell which answers were correct when presented with options 90% of the time. When Davis and his co-authors used the compound AI method, the layered model produced the correct answer 37% of the time — a nearly 10-fold increase.
Bootstrapping improvements to foundation models isn't exactly more efficient — this method is almost certainly a more expensive way to leverage AI models, but in some cases, the improved results are well worth the cost, Davis said.
He indicated that spending more on training and iterating now could save on inference computing in the future.
"I think people are getting more sophisticated at thinking about cost in more of a life-cycle way," Davis said.
Have you tested Chat GPT-5? Have a tip or an insight to share? Contact Emma at ecosgrove@businessinsider.com or use the secure messaging app Signal: 443-333-9088
